How Local AI Framing Shapes User Experience and Perceived Data Security with an Embodied AI Psychotherapist in VR

Abstract

The prevalence of anxiety disorders has been increasing in recent years to an extent where the demand for treatment is difficult to meet with traditional therapy. Therefore, research efforts aim for scalable, cost-effective extensions thereof. One approach is the use of Virtual Reality-based exposure therapy controlled by an embodied AI therapist. However, stakeholders, such as patients and therapists, are reluctant to embrace this development due to data security concerns. Along with the global rise of cloud-based large language models, smaller local models have improved to the point where they can simulate social interaction in a convincing manner. This leads to greater data security at the cost of a slightly decreased capability. In our study, we let N = 60 participants interact with an embodied conversational agent controlled by local models (i.e., Speech-to-Text, Large Language Model, Text-to-Speech). One group was informed that they were interacting with cloud-based models, while the other group was told they were interacting with locally hosted, on-premise models. Our results reveal significantly higher user experience and perceived quality of response for the ’local’ condition. Further analysis indicates a significant increase in perceived data security regarding general AI use for the ’local’ condition after the interaction and a slight decrease for the ’cloud-based’ condition.